open-agent-sdk-typescript

MEDIUM W011: Third-party content exposure detected (indirect prompt injection risk).

• Third-party content exposure detected (high risk: 0.90). The SKILL.md "In-Process MCP Servers" example defines an http_get tool that calls fetch(url) and shows the agent prompting "Fetch https://api.github.com/..." — meaning the agent can fetch and ingest arbitrary public URLs (untrusted third‑party content) that can influence its subsequent reasoning and tool use.

MEDIUM W012: Unverifiable external dependency detected (runtime URL that controls agent).

• Potentially malicious external URL detected (high risk: 0.90). The MCP server examples explicitly run npx to fetch and execute remote npm packages at runtime (e.g., "npx -y @modelcontextprotocol/server-filesystem" and "npx -y @modelcontextprotocol/server-postgres"), which will download and run external code the agent can depend on for MCP/tool functionality.

MEDIUM W013: Attempt to modify system services in skill instructions.

• Attempt to modify system services in skill instructions detected (medium risk: 0.60). The prompt exposes powerful state-modifying tools (Read/Write/Delete, Bash/BashSession, running npx/npm commands and MCP servers) and defaults to permissionMode "bypassPermissions", which enables an agent to execute shell commands and modify the filesystem or install software on the host (though it does not explicitly instruct obtaining sudo, creating users, or editing system-level configs like systemctl/ssh), so it poses a meaningful risk of compromising machine state.